知识卡片

普通读书笔记卡 · 9049.d · 生成模型:GPT-6

单目自监督保留尺度歧义

理解等级:高级

在无外部米制约束的单目重投影中,同时按s缩放预测深度和位姿平移,归一化投影保持不变。因此重建训练本身不提供绝对尺度;需已知基线、运动或真实尺寸等额外参照。低重建损失不能单独证明米制深度正确。

这张卡片看不懂?

已有反馈 · 0 条

    还没有人提交反馈。

    在整书笔记中的位置

    原书出处 · 1 处

    • · 查看原书依据

      The functions f1 and f2 are unknown and we want to learn them in an unsupervised way. If we have very long video sequences captured with a moving camera (assuming all objects are static) we can optimize the parameters of the functions f1 and f2 to minimize the reconstruction error. The functions f1 and f2 can be neural networks and their parameters can be obtained

    参考来源

    - 位置:《Foundations of Computer Vision (Antonio Torralba, Phillip Isola etc.)》第43章《单图深度估计》(来源项目:a19c7933-4eda-4176-aebd-30fe7e1ad292;源文件:extracted/Foundations of Computer Vision (Antonio Torralba, Phillip Isola etc.)/_epub-src/OPS/xhtml/c043.xhtml) - 结论依据:摘录中关于单目自监督保留尺度歧义的定义、比较或条件支撑卡片;原书表述不严谨的范围已在覆盖记录逐项限定,不把案例结果外推为普遍结论。 - 原始内容:The functions f1 and f2 are unknown and we want to learn them in an unsupervised way. If we have very long video sequences captured with a moving camera (assuming all objects are static) we can optimize the parameters of the functions f1 and f2 to minimize the reconstruction error. The functions f1 and f2 can be neural networks and their parameters can be obtained